Hello, everyone.
After some correspondence with Peter Geoghegan (1) and his ideas, I
have reworked the patch a lot and now it is much more simple with even
better performance (no new WAL or conflict resolution, hot standby
feedback is unrelated).
The idea is pretty simple now - let’s mark the page with
“standby-safe” LP_DEAD hints by the bit in btpo_flags
(BTP_LP_SAFE_ON_STANDBY and similar for gist and hash).
If standby wants to set LP_DEAD - it checks BTP_LP_SAFE_ON_STANDBY on
the page first, if it is not set - all “primary” hints are removed
first, and then the flag is set (with memory barrier to avoid memory
ordering issues in concurrent scans).
Also, standby checks BTP_LP_SAFE_ON_STANDBY to be sure about ignoring
tuples marked by LP_DEAD during the scan.
Of course, it is not so easy. If standby was promoted (or primary was
restored from standby backup) - it is still possible to receive FPI
with such flag set in WAL logs. So, the main problem is still there.
But we could just clear this flag while applying FPI because the page
remains dirty after that anyway! It should not cause any checksum,
consistency, or pg_rewind issues as explained in (2).
Semantically it is the same as set hint bit one milisecond after FPI
was applied (while page still remains dirty after FPI replay) - and
standby already does it with *heap* hint bits.
Also, TAP-test attached to (2) shows how it is easy to flush a hint
bit which was set by standby to achieve different checksum comparing
to primary already.
If standby was promoted (or restored from standby backup) it is safe
to use LP_DEAD with or without BTP_LP_SAFE_ON_STANDBY on a page. But
for accuracy BTP_LP_SAFE_ON_STANDBY is cleared by primary if found.
Also, we should take into account minRecoveryPoint as described in (3)
to avoid consistency issues during crash recovery (see
IsIndexLpDeadAllowed).
Also, as far as I know - there is no practical sense to keep
minRecoveryPoint at a low value. So, there is an optional patch that
moves minRecoveryPoint forward at each xl_running_data (to allow
standby to set hint bits and LP_DEADs more aggressively). It is about
every 15s.
There are some graphics showing performance testing results on my PC
in the attachment (test is taken from (4)). Each test was running for
10 minutes.
Additional primary performance is probably just measurement error. But
standby performance gain is huge.
Feel free to ask if you need more proof about correctness.
Thanks,
Michail.
[1] - https://www.postgresql.org/message-id/flat/CAH2-Wz%3D-BoaKgkN-MnKj6hFwO1BOJSA%2ByLMMO%2BLRZK932fNUXA%40mail.gmail.com#6d7cdebd68069cc493c11b9732fd2040
[2] - https://www.postgresql.org/message-id/flat/CANtu0oiAtteJ%2BMpPonBg6WfEsJCKrxuLK15P6GsaGDcYGjefVQ%40mail.gmail.com#091fca433185504f2818d5364819f7a4
[3] - https://www.postgresql.org/message-id/flat/CANtu0oh28mX5gy5jburH%2Bn1mcczK5_dCQnhbBnCM%3DPfqh-A26Q%40mail.gmail.com#ecfe5a331a3058f895c0cba698fbc4d3
[4] - https://www.postgresql.org/message-id/flat/CANtu0oiP18H31dSaEzn0B0rW6tA_q1G7%3D9Y92%2BUS_WHGOoQevg%40mail.gmail.com
